Advances in Depth Estimation and Visual Perception

The field of computer vision is rapidly advancing, with a focus on improving depth estimation and visual perception in various environments and conditions. Recent research has explored the use of monocular foundation models, LiDAR-visual-thermal datasets, and symmetry guidance for point cloud completion to achieve state-of-the-art performance. Notably, the integration of motion and structure priors has enabled robust depth estimation in diverse outdoor conditions. Furthermore, the development of benchmarks such as RGB-Th-Bench and LENVIZ has facilitated the evaluation of vision-language models and low-light image enhancement techniques. Innovative approaches, including intrinsic image decomposition and patchwise refinement, have also been proposed to address challenges in self-supervised monocular depth estimation and high-resolution image processing. Noteworthy papers include Distilling Monocular Foundation Model for Fine-grained Depth Completion, which achieved first place on the KITTI benchmark, and SymmCompletion, which introduced a highly effective point cloud completion method based on symmetry guidance. Additionally, DiffV2IR proposed a novel framework for visible-to-infrared image translation, and Synthetic-to-Real Self-supervised Robust Depth Estimation via Learning with Motion and Structure Priors presented a robust depth estimation framework incorporating motion and structure priors.

Sources

Distilling Monocular Foundation Model for Fine-grained Depth Completion

R-LiViT: A LiDAR-Visual-Thermal Dataset Enabling Vulnerable Road User Focused Roadside Perception

SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance

Distilling Stereo Networks for Performant and Efficient Leaner Networks

LeanStereo: A Leaner Backbone based Stereo Network

DiffV2IR: Visible-to-Infrared Diffusion Model via Vision-Language Understanding

RGB-Th-Bench: A Dense benchmark for Visual-Thermal Understanding of Vision Language Models

LENVIZ: A High-Resolution Low-Exposure Night Vision Benchmark Dataset

Synthetic-to-Real Self-supervised Robust Depth Estimation via Learning with Motion and Structure Priors

Deep Depth Estimation from Thermal Image: Dataset, Benchmark, and Challenges

Permutation-Invariant and Orientation-Aware Dataset Distillation for 3D Point Clouds

Intrinsic Image Decomposition for Robust Self-supervised Monocular Depth Estimation on Reflective Surfaces

One Look is Enough: A Novel Seamless Patchwise Refinement for Zero-Shot Monocular Depth Estimation Models on High-Resolution Images

Built with on top of